What we're learning building and running production AI—for our own products and for the businesses we work with. No hype, just the parts that actually matter.
85 articles from our team
Design agent tool permissions around bounded actions, server-side authorization, scoped credentials and tests that verify unauthorized actions stay blocked.
Design production language model monitoring that connects workflow traces, quality checks, cost and security signals to clear operational decisions.
Turn business records into a representative AI evaluation set, with sampling rules, reviewed answers, exception tests and measurable release criteria.
Design AI-assisted CRM updates with evidence-backed fields, safe record matching, conflict checks, exception handling and measurable release gates.
Build an email attachment workflow with reliable intake, bounded AI extraction, exception handling, approval controls and measurable release criteria.
Decide whether a local language model fits your regulated workflow using data boundaries, operating requirements, acceptance tests and a worked example.
Design approval gates that bind human decisions to exact AI-proposed actions, with clear ownership, expiry, exception handling and measurable tests.
Build document retrieval that enforces user access before generation, handles permission changes, and tests for leaks across answers, caches and citations.
Compare automation candidates using eligible volume, error consequences and exception handling costs, then choose a bounded pilot with measurable gates.